The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill describes how to compile and load machine learning models at runtime from computed paths using MLModel.compileModel(at:) and MLModel(contentsOf:). This capability allows for the execution of logic not contained in the original application bundle.
[REMOTE_CODE_EXECUTION]: The documentation outlines a workflow for downloading model files from remote URLs, followed by on-device compilation and execution.
Evidence: SKILL.md and references/profiling-and-personalization.md describe fetching a model from a downloadedModelURL, compiling it, and then loading the result.
[EXTERNAL_DOWNLOADS]: The skill contains instructions for fetching machine learning models from remote network locations at runtime.
[COMMAND_EXECUTION]: The skill describes performing file system operations, such as moving compiled model directories and persisting fine-tuned models to the application support directory using FileManager and MLWritable.write(to:).
[INDIRECT_PROMPT_INJECTION]: The skill identifies surfaces for processing untrusted external data, such as images and labels, which could potentially influence application behavior if the model is susceptible to malicious inputs.
Ingestion points: VNCoreMLRequest inputs and MLUpdateTask training data.
Boundary markers: None specified.
Capability inventory: File system access and runtime model compilation.
Sanitization: The skill recommends performing integrity checks (signature/hash validation) for downloaded models before compilation.

coreml